$ the-wire · showcase
vLLM 0.28.0 LANDED WITH 584 COMMITS AND A KIMI-K3 PERFORMANCE PUSH
By RepoJournal · Filed · About Local LLMs · Composed from the cited sources · methodology
14 people shipped this
ParthSareen
@ParthSareen
2 cited
jessegross
@jessegross
1 cited
Dominik Pantaleoni
@dpantaleoni
1 cited
khluu
@khluu
1 cited
VBS2004
@VBS2004
1 cited
JustinTong0323
@JustinTong0323
1 cited
hnyls2002
@hnyls2002
1 cited
hoyyeva
@hoyyeva
Aleksander Grygier
@allozaur
Max Krasnyansky
@max-krasnyansky
oliverholworthy
@oliverholworthy
Woosuk Kwon
@WoosukKwon
Fangzhou-Ai
@Fangzhou-Ai
ch-wan
@ch-wan
vLLM shipped its biggest release in months, and Ollama quietly fixed a macOS bug that could leave stale app processes running.
vLLM's v0.28.0 release is the big one this cycle, with 584 commits from 270 contributors and a major Kimi-K3 optimization effort spanning Decode Context Parallel and fused FlashKDA kernels [1]. Meanwhile, two Ollama PRs from ParthSareen tackle macOS app process handoff and a proxy race condition that caused intermittent "Claude model catalog changed" errors [2][3]. The handoff change is particularly important: it favors availability, but if the election fails, a stale instance may remain, so test your upgrade path [2]. On the model side, Ollama is removing the MLX gemma3 text-only port because it lacks vision support and could misroute requests [4]. llama.cpp also pushed a chunked SSD MMA optimization for Mamba-2 prefill on Metal, promising parallel matmuls for token sequences that fit 64-token chunks [5]. In sglang, they added day-0 support for Ling-3.0-flash, a hybrid MoE with KDA linear attention and MLA, plus a new beam search feature that runs k-wide beams without a server flag [6][7]. Finally, vLLM fixed a bug where unguarded tool call JSON parsing could take down conversations permanently [8].
Action items
- → Upgrade vLLM to v0.28.0 for the Kimi-K3 performance gains and the chat postprocessing bugfix vllm-project/vllm [immediate]
- → Review the Ollama macOS handoff changes before rolling out to production Macs ollama/ollama [plan]
- → Test sglang's new beam search feature in a staging environment before relying on it sgl-project/sglang [monitor]
References
- [1] v0.28.0 ↗ vllm-project/vllm
- [2] app: synchronize macOS app handoff ↗ ollama/ollama
- [3] proxy: continue requests when the model catalog changes ↗ ollama/ollama
- [4] mlx: remove the text-only gemma3 model ↗ ollama/ollama
- [5] ggml-metal: add chunked SSD MMA for Mamba-2 prefill optimization ↗ ggml-org/llama.cpp
- [6] [Model] Support Ling-3.0-flash (BailingMoeV3) ↗ sgl-project/sglang
- [7] [Feature] Beam search support ↗ sgl-project/sglang
- [8] [Bugfix] Guard tool call argument JSON parsing in chat message postprocessing ↗ vllm-project/vllm
